Classification of SAR imagery based on the multiscale stochastic process

This paper presents efficient non-parameter and multiscale approach to segmentation of natural clutter in synthetic aperture radar (SAR) imagery. The method we propose not only exploit the coherent nature of SAR sensor, and take of advantage of the characteristic statistical difference in imagery of different terrain types, but also do not require distribution of pixels due to using bootstrap method. Firstly, we employ multiscale autoregressive (MAR) model for describing random processes that evolve in scale. Secondly, the confidence internals of the parameters in MAR model for each category of terrain of interest are calculated using bootstrap technical. Then, for each pixel, we generate a set of parameter estimation that characterizes the local evolution in scale. The pixel is classified by relation between the parameters estimation and the confidence internal of each type. Finally, test images are classified to demonstrate the method.